[X86] Move RDFLAGS/WRFLAGS expansion until after RA

[void]

Summary

The register allocator may introduce reloads in the middle of reading
and writing the EFLAGS register, due to the RDFLAGS & WRFLAGS pseudos
being expanded before RA. This may cause an issue where the stack
pointer was adjusted but the stack offset for the reload wasn't
accounted for (see [1]).

To avoid this, expand these pseudos after register allocation.

[1] https://github.com/llvm/llvm-project/issues/59102

[arsenm]

llvm/lib/Target/X86/X86InstrInfo.cpp

5074–5075

Are these not reserved?

[void]

llvm/lib/Target/X86/X86InstrInfo.cpp

5074–5075

Dunno. I'm going off of what the original code was doing...

[nickdesaulniers]

llvm/test/CodeGen/X86/eflags.ll

4–22 ↗

(On Diff #482958)

These are already covered by llvm/test/CodeGen/X86/x86-64-flags-intrinsics.ll. Do they add anything here?

[efriedma]

You should be able to teach frame lowering to compute the correct offsets by fixing X86InstrInfo::getSPAdjust(). Maybe that's better than trying to avoid the use of frame indexes between the push and pop?

[pengfei]

llvm/lib/Target/X86/X86InstrInfo.td

1384

Seems not needed.

1397

ditto.

[craig.topper]

llvm/lib/Target/X86/X86InstrInfo.td

1384

I think it's need to make expandPostRAPseudo get called.

[pengfei]

llvm/lib/Target/X86/X86InstrInfo.td

1384

It just checks isPseudo https://github.com/llvm/llvm-project/blob/main/llvm/lib/CodeGen/ExpandPostRAPseudos.cpp#L199-L203

[pengfei]

llvm/lib/Target/X86/X86InstrInfo.td

1384

I mean usesCustomInserter = 1 is not needed since the code is removed from EmitInstrWithCustomInserter.

[craig.topper]

llvm/lib/Target/X86/X86InstrInfo.td

1384

Agreed. usesCustomInserter can be removed

[pengfei]

llvm/lib/Target/X86/X86InstrInfo.cpp

5074–5075

I think this might be a problem for PostRA.

[pengfei]

llvm/lib/Target/X86/X86InstrInfo.cpp

5074–5075

Sorry, I misunderstood the question. Please ignore.

[void]

In D140045#3996240, @efriedma wrote:

You should be able to teach frame lowering to compute the correct offsets by fixing X86InstrInfo::getSPAdjust(). Maybe that's better than trying to avoid the use of frame indexes between the push and pop?

That's the correct general fix for this issue. Though I still maintain that this is a good addition. Having reloads occurring in the middle of these two instructions isn't really beneficial.

To be clear, I'll submit a subsequent patch that corrects the offsets. :-)

llvm/lib/Target/X86/X86InstrInfo.td

1384

Ah! Okay.

[void]

Update TD file and testcase. Also correct the MIBs to have the correct parameters.

[craig.topper]

LGTM

[pengfei]

LGTM.

[void]

Rebase.